2 * address space "slices" (meta-segments) support
4 * Copyright (C) 2007 Benjamin Herrenschmidt, IBM Corporation.
6 * Based on hugetlb implementation
8 * Copyright (C) 2003 David Gibson, IBM Corporation.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 #include <linux/kernel.h>
29 #include <linux/pagemap.h>
30 #include <linux/err.h>
31 #include <linux/spinlock.h>
32 #include <linux/export.h>
33 #include <linux/hugetlb.h>
36 #include <asm/copro.h>
37 #include <asm/hugetlb.h>
39 static DEFINE_SPINLOCK(slice_convert_lock);
44 static void slice_print_mask(const char *label, const struct slice_mask *mask)
48 pr_devel("%s low_slice: %*pbl\n", label,
49 (int)SLICE_NUM_LOW, &mask->low_slices);
50 pr_devel("%s high_slice: %*pbl\n", label,
51 (int)SLICE_NUM_HIGH, mask->high_slices);
54 #define slice_dbg(fmt...) do { if (_slice_debug) pr_devel(fmt); } while (0)
58 static void slice_print_mask(const char *label, const struct slice_mask *mask) {}
59 #define slice_dbg(fmt...)
63 static void slice_range_to_mask(unsigned long start, unsigned long len,
64 struct slice_mask *ret)
66 unsigned long end = start + len - 1;
70 bitmap_zero(ret->high_slices, SLICE_NUM_HIGH);
72 if (start < SLICE_LOW_TOP) {
73 unsigned long mend = min(end,
74 (unsigned long)(SLICE_LOW_TOP - 1));
76 ret->low_slices = (1u << (GET_LOW_SLICE_INDEX(mend) + 1))
77 - (1u << GET_LOW_SLICE_INDEX(start));
80 if ((start + len) > SLICE_LOW_TOP) {
81 unsigned long start_index = GET_HIGH_SLICE_INDEX(start);
82 unsigned long align_end = ALIGN(end, (1UL << SLICE_HIGH_SHIFT));
83 unsigned long count = GET_HIGH_SLICE_INDEX(align_end) - start_index;
85 bitmap_set(ret->high_slices, start_index, count);
89 static int slice_area_is_free(struct mm_struct *mm, unsigned long addr,
92 struct vm_area_struct *vma;
94 if ((mm->context.slb_addr_limit - len) < addr)
96 vma = find_vma(mm, addr);
97 return (!vma || (addr + len) <= vm_start_gap(vma));
100 static int slice_low_has_vma(struct mm_struct *mm, unsigned long slice)
102 return !slice_area_is_free(mm, slice << SLICE_LOW_SHIFT,
103 1ul << SLICE_LOW_SHIFT);
106 static int slice_high_has_vma(struct mm_struct *mm, unsigned long slice)
108 unsigned long start = slice << SLICE_HIGH_SHIFT;
109 unsigned long end = start + (1ul << SLICE_HIGH_SHIFT);
112 /* Hack, so that each addresses is controlled by exactly one
113 * of the high or low area bitmaps, the first high area starts
116 start = SLICE_LOW_TOP;
119 return !slice_area_is_free(mm, start, end - start);
122 static void slice_mask_for_free(struct mm_struct *mm, struct slice_mask *ret,
123 unsigned long high_limit)
129 bitmap_zero(ret->high_slices, SLICE_NUM_HIGH);
131 for (i = 0; i < SLICE_NUM_LOW; i++)
132 if (!slice_low_has_vma(mm, i))
133 ret->low_slices |= 1u << i;
135 if (high_limit <= SLICE_LOW_TOP)
138 for (i = 0; i < GET_HIGH_SLICE_INDEX(high_limit); i++)
139 if (!slice_high_has_vma(mm, i))
140 __set_bit(i, ret->high_slices);
143 #ifdef CONFIG_PPC_BOOK3S_64
144 static struct slice_mask *slice_mask_for_size(struct mm_struct *mm, int psize)
146 #ifdef CONFIG_PPC_64K_PAGES
147 if (psize == MMU_PAGE_64K)
148 return &mm->context.mask_64k;
150 if (psize == MMU_PAGE_4K)
151 return &mm->context.mask_4k;
152 #ifdef CONFIG_HUGETLB_PAGE
153 if (psize == MMU_PAGE_16M)
154 return &mm->context.mask_16m;
155 if (psize == MMU_PAGE_16G)
156 return &mm->context.mask_16g;
160 #elif defined(CONFIG_PPC_8xx)
161 static struct slice_mask *slice_mask_for_size(struct mm_struct *mm, int psize)
163 if (psize == mmu_virtual_psize)
164 return &mm->context.mask_base_psize;
165 #ifdef CONFIG_HUGETLB_PAGE
166 if (psize == MMU_PAGE_512K)
167 return &mm->context.mask_512k;
168 if (psize == MMU_PAGE_8M)
169 return &mm->context.mask_8m;
174 #error "Must define the slice masks for page sizes supported by the platform"
177 static bool slice_check_range_fits(struct mm_struct *mm,
178 const struct slice_mask *available,
179 unsigned long start, unsigned long len)
181 unsigned long end = start + len - 1;
184 if (start < SLICE_LOW_TOP) {
185 unsigned long mend = min(end,
186 (unsigned long)(SLICE_LOW_TOP - 1));
188 low_slices = (1u << (GET_LOW_SLICE_INDEX(mend) + 1))
189 - (1u << GET_LOW_SLICE_INDEX(start));
191 if ((low_slices & available->low_slices) != low_slices)
194 if (SLICE_NUM_HIGH && ((start + len) > SLICE_LOW_TOP)) {
195 unsigned long start_index = GET_HIGH_SLICE_INDEX(start);
196 unsigned long align_end = ALIGN(end, (1UL << SLICE_HIGH_SHIFT));
197 unsigned long count = GET_HIGH_SLICE_INDEX(align_end) - start_index;
200 for (i = start_index; i < start_index + count; i++) {
201 if (!test_bit(i, available->high_slices))
209 static void slice_flush_segments(void *parm)
212 struct mm_struct *mm = parm;
215 if (mm != current->active_mm)
218 copy_mm_to_paca(current->active_mm);
220 local_irq_save(flags);
221 slb_flush_and_rebolt();
222 local_irq_restore(flags);
226 static void slice_convert(struct mm_struct *mm,
227 const struct slice_mask *mask, int psize)
229 int index, mask_index;
230 /* Write the new slice psize bits */
231 unsigned char *hpsizes, *lpsizes;
232 struct slice_mask *psize_mask, *old_mask;
233 unsigned long i, flags;
236 slice_dbg("slice_convert(mm=%p, psize=%d)\n", mm, psize);
237 slice_print_mask(" mask", mask);
239 psize_mask = slice_mask_for_size(mm, psize);
241 /* We need to use a spinlock here to protect against
242 * concurrent 64k -> 4k demotion ...
244 spin_lock_irqsave(&slice_convert_lock, flags);
246 lpsizes = mm->context.low_slices_psize;
247 for (i = 0; i < SLICE_NUM_LOW; i++) {
248 if (!(mask->low_slices & (1u << i)))
251 mask_index = i & 0x1;
254 /* Update the slice_mask */
255 old_psize = (lpsizes[index] >> (mask_index * 4)) & 0xf;
256 old_mask = slice_mask_for_size(mm, old_psize);
257 old_mask->low_slices &= ~(1u << i);
258 psize_mask->low_slices |= 1u << i;
260 /* Update the sizes array */
261 lpsizes[index] = (lpsizes[index] & ~(0xf << (mask_index * 4))) |
262 (((unsigned long)psize) << (mask_index * 4));
265 hpsizes = mm->context.high_slices_psize;
266 for (i = 0; i < GET_HIGH_SLICE_INDEX(mm->context.slb_addr_limit); i++) {
267 if (!test_bit(i, mask->high_slices))
270 mask_index = i & 0x1;
273 /* Update the slice_mask */
274 old_psize = (hpsizes[index] >> (mask_index * 4)) & 0xf;
275 old_mask = slice_mask_for_size(mm, old_psize);
276 __clear_bit(i, old_mask->high_slices);
277 __set_bit(i, psize_mask->high_slices);
279 /* Update the sizes array */
280 hpsizes[index] = (hpsizes[index] & ~(0xf << (mask_index * 4))) |
281 (((unsigned long)psize) << (mask_index * 4));
284 slice_dbg(" lsps=%lx, hsps=%lx\n",
285 (unsigned long)mm->context.low_slices_psize,
286 (unsigned long)mm->context.high_slices_psize);
288 spin_unlock_irqrestore(&slice_convert_lock, flags);
290 copro_flush_all_slbs(mm);
294 * Compute which slice addr is part of;
295 * set *boundary_addr to the start or end boundary of that slice
296 * (depending on 'end' parameter);
297 * return boolean indicating if the slice is marked as available in the
298 * 'available' slice_mark.
300 static bool slice_scan_available(unsigned long addr,
301 const struct slice_mask *available,
302 int end, unsigned long *boundary_addr)
305 if (addr < SLICE_LOW_TOP) {
306 slice = GET_LOW_SLICE_INDEX(addr);
307 *boundary_addr = (slice + end) << SLICE_LOW_SHIFT;
308 return !!(available->low_slices & (1u << slice));
310 slice = GET_HIGH_SLICE_INDEX(addr);
311 *boundary_addr = (slice + end) ?
312 ((slice + end) << SLICE_HIGH_SHIFT) : SLICE_LOW_TOP;
313 return !!test_bit(slice, available->high_slices);
317 static unsigned long slice_find_area_bottomup(struct mm_struct *mm,
319 const struct slice_mask *available,
320 int psize, unsigned long high_limit)
322 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
323 unsigned long addr, found, next_end;
324 struct vm_unmapped_area_info info;
328 info.align_mask = PAGE_MASK & ((1ul << pshift) - 1);
329 info.align_offset = 0;
331 addr = TASK_UNMAPPED_BASE;
333 * Check till the allow max value for this mmap request
335 while (addr < high_limit) {
336 info.low_limit = addr;
337 if (!slice_scan_available(addr, available, 1, &addr))
342 * At this point [info.low_limit; addr) covers
343 * available slices only and ends at a slice boundary.
344 * Check if we need to reduce the range, or if we can
345 * extend it to cover the next available slice.
347 if (addr >= high_limit)
349 else if (slice_scan_available(addr, available, 1, &next_end)) {
353 info.high_limit = addr;
355 found = vm_unmapped_area(&info);
356 if (!(found & ~PAGE_MASK))
363 static unsigned long slice_find_area_topdown(struct mm_struct *mm,
365 const struct slice_mask *available,
366 int psize, unsigned long high_limit)
368 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
369 unsigned long addr, found, prev;
370 struct vm_unmapped_area_info info;
372 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
374 info.align_mask = PAGE_MASK & ((1ul << pshift) - 1);
375 info.align_offset = 0;
377 addr = mm->mmap_base;
379 * If we are trying to allocate above DEFAULT_MAP_WINDOW
380 * Add the different to the mmap_base.
381 * Only for that request for which high_limit is above
382 * DEFAULT_MAP_WINDOW we should apply this.
384 if (high_limit > DEFAULT_MAP_WINDOW)
385 addr += mm->context.slb_addr_limit - DEFAULT_MAP_WINDOW;
387 while (addr > PAGE_SIZE) {
388 info.high_limit = addr;
389 if (!slice_scan_available(addr - 1, available, 0, &addr))
394 * At this point [addr; info.high_limit) covers
395 * available slices only and starts at a slice boundary.
396 * Check if we need to reduce the range, or if we can
397 * extend it to cover the previous available slice.
399 if (addr < PAGE_SIZE)
401 else if (slice_scan_available(addr - 1, available, 0, &prev)) {
405 info.low_limit = addr;
407 found = vm_unmapped_area(&info);
408 if (!(found & ~PAGE_MASK))
413 * A failed mmap() very likely causes application failure,
414 * so fall back to the bottom-up function here. This scenario
415 * can happen with large stack limits and large mmap()
418 return slice_find_area_bottomup(mm, len, available, psize, high_limit);
422 static unsigned long slice_find_area(struct mm_struct *mm, unsigned long len,
423 const struct slice_mask *mask, int psize,
424 int topdown, unsigned long high_limit)
427 return slice_find_area_topdown(mm, len, mask, psize, high_limit);
429 return slice_find_area_bottomup(mm, len, mask, psize, high_limit);
432 static inline void slice_copy_mask(struct slice_mask *dst,
433 const struct slice_mask *src)
435 dst->low_slices = src->low_slices;
438 bitmap_copy(dst->high_slices, src->high_slices, SLICE_NUM_HIGH);
441 static inline void slice_or_mask(struct slice_mask *dst,
442 const struct slice_mask *src1,
443 const struct slice_mask *src2)
445 dst->low_slices = src1->low_slices | src2->low_slices;
448 bitmap_or(dst->high_slices, src1->high_slices, src2->high_slices, SLICE_NUM_HIGH);
451 static inline void slice_andnot_mask(struct slice_mask *dst,
452 const struct slice_mask *src1,
453 const struct slice_mask *src2)
455 dst->low_slices = src1->low_slices & ~src2->low_slices;
458 bitmap_andnot(dst->high_slices, src1->high_slices, src2->high_slices, SLICE_NUM_HIGH);
461 #ifdef CONFIG_PPC_64K_PAGES
462 #define MMU_PAGE_BASE MMU_PAGE_64K
464 #define MMU_PAGE_BASE MMU_PAGE_4K
467 unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len,
468 unsigned long flags, unsigned int psize,
471 struct slice_mask good_mask;
472 struct slice_mask potential_mask;
473 const struct slice_mask *maskp;
474 const struct slice_mask *compat_maskp = NULL;
475 int fixed = (flags & MAP_FIXED);
476 int pshift = max_t(int, mmu_psize_defs[psize].shift, PAGE_SHIFT);
477 unsigned long page_size = 1UL << pshift;
478 struct mm_struct *mm = current->mm;
479 unsigned long newaddr;
480 unsigned long high_limit;
482 high_limit = DEFAULT_MAP_WINDOW;
483 if (addr >= high_limit || (fixed && (addr + len > high_limit)))
484 high_limit = TASK_SIZE;
486 if (len > high_limit)
488 if (len & (page_size - 1))
491 if (addr & (page_size - 1))
493 if (addr > high_limit - len)
497 if (high_limit > mm->context.slb_addr_limit) {
499 * Increasing the slb_addr_limit does not require
500 * slice mask cache to be recalculated because it should
501 * be already initialised beyond the old address limit.
503 mm->context.slb_addr_limit = high_limit;
505 on_each_cpu(slice_flush_segments, mm, 1);
509 BUG_ON(mm->task_size == 0);
510 BUG_ON(mm->context.slb_addr_limit == 0);
511 VM_BUG_ON(radix_enabled());
513 slice_dbg("slice_get_unmapped_area(mm=%p, psize=%d...\n", mm, psize);
514 slice_dbg(" addr=%lx, len=%lx, flags=%lx, topdown=%d\n",
515 addr, len, flags, topdown);
517 /* If hint, make sure it matches our alignment restrictions */
518 if (!fixed && addr) {
519 addr = _ALIGN_UP(addr, page_size);
520 slice_dbg(" aligned addr=%lx\n", addr);
521 /* Ignore hint if it's too large or overlaps a VMA */
522 if (addr > high_limit - len ||
523 !slice_area_is_free(mm, addr, len))
527 /* First make up a "good" mask of slices that have the right size
530 maskp = slice_mask_for_size(mm, psize);
533 * Here "good" means slices that are already the right page size,
534 * "compat" means slices that have a compatible page size (i.e.
535 * 4k in a 64k pagesize kernel), and "free" means slices without
539 * check if fits in good | compat => OK
540 * check if fits in good | compat | free => convert free
543 * check if hint fits in good => OK
544 * check if hint fits in good | free => convert free
546 * search in good, found => OK
547 * search in good | free, found => convert free
548 * search in good | compat | free, found => convert free.
552 * If we support combo pages, we can allow 64k pages in 4k slices
553 * The mask copies could be avoided in most cases here if we had
554 * a pointer to good mask for the next code to use.
556 if (IS_ENABLED(CONFIG_PPC_64K_PAGES) && psize == MMU_PAGE_64K) {
557 compat_maskp = slice_mask_for_size(mm, MMU_PAGE_4K);
559 slice_or_mask(&good_mask, maskp, compat_maskp);
561 slice_copy_mask(&good_mask, maskp);
563 slice_copy_mask(&good_mask, maskp);
566 slice_print_mask(" good_mask", &good_mask);
568 slice_print_mask(" compat_mask", compat_maskp);
570 /* First check hint if it's valid or if we have MAP_FIXED */
571 if (addr != 0 || fixed) {
572 /* Check if we fit in the good mask. If we do, we just return,
575 if (slice_check_range_fits(mm, &good_mask, addr, len)) {
576 slice_dbg(" fits good !\n");
580 /* Now let's see if we can find something in the existing
581 * slices for that size
583 newaddr = slice_find_area(mm, len, &good_mask,
584 psize, topdown, high_limit);
585 if (newaddr != -ENOMEM) {
586 /* Found within the good mask, we don't have to setup,
587 * we thus return directly
589 slice_dbg(" found area at 0x%lx\n", newaddr);
594 * We don't fit in the good mask, check what other slices are
595 * empty and thus can be converted
597 slice_mask_for_free(mm, &potential_mask, high_limit);
598 slice_or_mask(&potential_mask, &potential_mask, &good_mask);
599 slice_print_mask(" potential", &potential_mask);
601 if (addr != 0 || fixed) {
602 if (slice_check_range_fits(mm, &potential_mask, addr, len)) {
603 slice_dbg(" fits potential !\n");
608 /* If we have MAP_FIXED and failed the above steps, then error out */
612 slice_dbg(" search...\n");
614 /* If we had a hint that didn't work out, see if we can fit
615 * anywhere in the good area.
618 addr = slice_find_area(mm, len, &good_mask,
619 psize, topdown, high_limit);
620 if (addr != -ENOMEM) {
621 slice_dbg(" found area at 0x%lx\n", addr);
626 /* Now let's see if we can find something in the existing slices
627 * for that size plus free slices
629 addr = slice_find_area(mm, len, &potential_mask,
630 psize, topdown, high_limit);
632 #ifdef CONFIG_PPC_64K_PAGES
633 if (addr == -ENOMEM && psize == MMU_PAGE_64K) {
634 /* retry the search with 4k-page slices included */
635 slice_or_mask(&potential_mask, &potential_mask, compat_maskp);
636 addr = slice_find_area(mm, len, &potential_mask,
637 psize, topdown, high_limit);
644 slice_range_to_mask(addr, len, &potential_mask);
645 slice_dbg(" found potential area at 0x%lx\n", addr);
646 slice_print_mask(" mask", &potential_mask);
649 slice_andnot_mask(&potential_mask, &potential_mask, &good_mask);
650 if (compat_maskp && !fixed)
651 slice_andnot_mask(&potential_mask, &potential_mask, compat_maskp);
652 if (potential_mask.low_slices ||
654 !bitmap_empty(potential_mask.high_slices, SLICE_NUM_HIGH))) {
655 slice_convert(mm, &potential_mask, psize);
656 if (psize > MMU_PAGE_BASE)
657 on_each_cpu(slice_flush_segments, mm, 1);
662 EXPORT_SYMBOL_GPL(slice_get_unmapped_area);
664 unsigned long arch_get_unmapped_area(struct file *filp,
670 return slice_get_unmapped_area(addr, len, flags,
671 current->mm->context.user_psize, 0);
674 unsigned long arch_get_unmapped_area_topdown(struct file *filp,
675 const unsigned long addr0,
676 const unsigned long len,
677 const unsigned long pgoff,
678 const unsigned long flags)
680 return slice_get_unmapped_area(addr0, len, flags,
681 current->mm->context.user_psize, 1);
684 unsigned int get_slice_psize(struct mm_struct *mm, unsigned long addr)
686 unsigned char *psizes;
687 int index, mask_index;
690 * Radix doesn't use slice, but can get enabled along with MMU_SLICE
692 if (radix_enabled()) {
693 #ifdef CONFIG_PPC_64K_PAGES
699 if (addr < SLICE_LOW_TOP) {
700 psizes = mm->context.low_slices_psize;
701 index = GET_LOW_SLICE_INDEX(addr);
703 psizes = mm->context.high_slices_psize;
704 index = GET_HIGH_SLICE_INDEX(addr);
706 mask_index = index & 0x1;
707 return (psizes[index >> 1] >> (mask_index * 4)) & 0xf;
709 EXPORT_SYMBOL_GPL(get_slice_psize);
711 void slice_init_new_context_exec(struct mm_struct *mm)
713 unsigned char *hpsizes, *lpsizes;
714 struct slice_mask *mask;
715 unsigned int psize = mmu_virtual_psize;
717 slice_dbg("slice_init_new_context_exec(mm=%p)\n", mm);
720 * In the case of exec, use the default limit. In the
721 * case of fork it is just inherited from the mm being
725 mm->context.slb_addr_limit = DEFAULT_MAP_WINDOW_USER64;
727 mm->context.slb_addr_limit = DEFAULT_MAP_WINDOW;
730 mm->context.user_psize = psize;
733 * Set all slice psizes to the default.
735 lpsizes = mm->context.low_slices_psize;
736 memset(lpsizes, (psize << 4) | psize, SLICE_NUM_LOW >> 1);
738 hpsizes = mm->context.high_slices_psize;
739 memset(hpsizes, (psize << 4) | psize, SLICE_NUM_HIGH >> 1);
742 * Slice mask cache starts zeroed, fill the default size cache.
744 mask = slice_mask_for_size(mm, psize);
745 mask->low_slices = ~0UL;
747 bitmap_fill(mask->high_slices, SLICE_NUM_HIGH);
750 void slice_set_range_psize(struct mm_struct *mm, unsigned long start,
751 unsigned long len, unsigned int psize)
753 struct slice_mask mask;
755 VM_BUG_ON(radix_enabled());
757 slice_range_to_mask(start, len, &mask);
758 slice_convert(mm, &mask, psize);
761 #ifdef CONFIG_HUGETLB_PAGE
763 * is_hugepage_only_range() is used by generic code to verify whether
764 * a normal mmap mapping (non hugetlbfs) is valid on a given area.
766 * until the generic code provides a more generic hook and/or starts
767 * calling arch get_unmapped_area for MAP_FIXED (which our implementation
768 * here knows how to deal with), we hijack it to keep standard mappings
771 * because of that generic code limitation, MAP_FIXED mapping cannot
772 * "convert" back a slice with no VMAs to the standard page size, only
773 * get_unmapped_area() can. It would be possible to fix it here but I
774 * prefer working on fixing the generic code instead.
776 * WARNING: This will not work if hugetlbfs isn't enabled since the
777 * generic code will redefine that function as 0 in that. This is ok
778 * for now as we only use slices with hugetlbfs enabled. This should
779 * be fixed as the generic code gets fixed.
781 int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr,
784 const struct slice_mask *maskp;
785 unsigned int psize = mm->context.user_psize;
790 maskp = slice_mask_for_size(mm, psize);
791 #ifdef CONFIG_PPC_64K_PAGES
792 /* We need to account for 4k slices too */
793 if (psize == MMU_PAGE_64K) {
794 const struct slice_mask *compat_maskp;
795 struct slice_mask available;
797 compat_maskp = slice_mask_for_size(mm, MMU_PAGE_4K);
798 slice_or_mask(&available, maskp, compat_maskp);
799 return !slice_check_range_fits(mm, &available, addr, len);
803 return !slice_check_range_fits(mm, maskp, addr, len);